MACHINE LEARNING AND BIOIMPEDANCE TO DRIVE CLINICAL DECISION-MAKING AT HOME: A PROPOSED MODEL TO PREDICT HEART FAILURE EXACERBATIONSAUTHORS

Abstract

Therapeutic Area: Heart Failure Background: Heart failure (HF) exacerbations lead to an estimated 1 million congestive heart failure (CHF)-related admissions annually, costing between $11-15,667 with a hospital stay of 7-21 days. The etiology is multifactorial, but regular monitoring is essential for preventing readmissions. We propose an at-home machine learning (ML) monitoring model, incorporating multiple variables and including bioimpedance technology to measure Extracellular fluid/total body water (ECF/TBW), to develop a predictive CHF exacerbation model that improves CHF care and reduces its burden on health systems. Methods: Trials investigating remote CHF monitoring have shown varying results. The TIM-HF2 trial showed a significant reduction in hospitalization and mortality when monitoring vitals, reporting symptoms, and conducting monthly telehealth visits. Another trial utilized ML models incorporating vitals, medication adherence, symptoms, demographics, diagnoses, and risk factors to triage patients better than physicians.No ML trials have incorporated ECF/TBW, despite its established correlation with inferior vena cava size on ultrasound, a surrogate marker of fluid overload status (Pearson coefficient=-0.73, p<0.0001). Simple 4-lead home smart weight scales are common and able to measure ECF/TBW. ECF/TBW is currently used mainly in clinical practice to guide fluid management. Results: We propose a remote CHF monitoring system utilizing a ML model incorporating demographics, risk factors, medication adherence, diet compliance, H2FPEF score, left-ventricular ejection fraction, and ECF/TBW measured by home smart-scales to predict CHF exacerbations. It would then triage patients into 3 categories: continue current plan, improve compliance and follow-up earlier, or seek emergent care. Conclusions: There is a need to predict CHF exacerbations before they escalate to hospitalization. Accurate prediction and earlier intervention may reduce associated health risks, costs, and mortality. No prior trial utilized bioimpedance-measured ECF/TBW, and ML-based CHF monitoring has strong preliminary results. Our proposed solution combines both approaches to better predict CHF exacerbation and triage patients to appropriate interventions. Next steps include collecting single-center CHF patient data and retrofitting an intrahospital HF monitoring ML model to include ECF/TBW, followed by a pilot prospective cohort study.